Enhancing Scholarly Paper Recommendation by Modelling Diversity of Research Interests

Recommender systems help researchers identify relevant papers in scientific document collections. A precise user interest model is crucial for content-based scholarly paper recommendation. Arguably, past publications play an important role in modelling researchers’ interests. However, not all publications account for the interest model equally. Existing approaches introduce weighting schemes to emphasize the impact of recent articles published by each researcher. However, these weighting schemes fail to explain the content-wise relationship (e.g. diversity) among their publications. In this paper, we introduce a new feature to capture the diversity of research interests derived from each researcher’s publications, which can be combined with such weighting schemes. We further employ this feature in two weighting schemes to model research interests for each researcher. We investigate the effect of the new feature with two text representation models to represent papers and compare the effectiveness of four weighting schemes to model user interest. We conduct experiments on a public dataset of 50 researchers. Results show that although the accuracy obtained with our proposed weighting schemes is not stable with different parameter settings, our methods in optimal settings reveal an increase in accuracy measured by NDCG@10 and P@10, compared to other existing weighting schemes.